Background
When do people visit the doctor? To answer this question, an important one for health policy planning, researchers analyze characteristics of the population under study (potential patients), medical institutions, and social structure as they relate to health care issues [
1]. It is necessary to categorize doctors by specialization – there are characteristics of a population whose variations can influence, in opposite directions, the probability of visiting General Practitioner (GP) against Specialist Doctor (SD). For example, in studying migrants’ health, authors cited income as a characteristic positively affecting visits to GP, and negatively affecting visits to SD [
2].
Preventive health behavior (PHB), or “activity undertaken by a person who believes himself to be healthy for the purpose of preventing disease” [
3], is one of the characteristics of a population which affects the frequency of visits to a doctor. Pender suggested that PHB and health-promoting behavior (e.g. physical activities) complement one another [
4], and the term PHB is used in this study to refer to these two interrelated dimensions of a healthy lifestyle. A number of recent studies report the influence of various PHB factors on medical care utilization including smoking [
5], obesity [
6], influenza vaccination [
7], physical activity [
8] or combination of PHB factors [
9]. Similar studies were undertaken in Israel: on smoking [
10], on physical activity [
11], on influenza vaccination [
12], and on mammography [
13].
The aim of this study is to examine the joint impact of PHB and social and demographic factors of individuals on the utilization of primary and secondary medical care, as measured by visits to both GP and SD doctors. In the case of Israel, it is assumed that due to universal health insurance, and the fact that more than half GPs are not paid through a fee-for-service system, the frequency of visits to the doctor are not greatly affected by variations in medical institutions and social structure.
Analysis of disadvantaged populations and their social-demographic characteristics is important for this study. For these populations, motives for engaging in PHB may be weaker due to low socio-economic status (SES) and lower income [
14,
15], advanced age [
16], gender [
17], residence in a disadvantaged neighborhood [
18], or lack of education [
19]. This weaker motivation can be explained by lack of social support [
20], lack of access to facilities for engaging in physical activity or a healthy diet [
21], or lack of awareness and understanding health care issues [
22]. A similar relationship is shown by many Israeli studies: for older populations [
23], for residents in the periphery [
24], for women [
25], and for people with lower SES [
26].
The problem of the inequity in PHB is reported for several countries with universal health insurance: in European countries [
16], Australia [
27], and Japan [
28].
In the current study, we treat “visits to the doctor” as a discrete ordered variable and utilize an ordered probit model to measure the joint impact and marginal effects of PHB and disadvantage factors on frequency of visits to the doctor. Concurrently, PHB can be itself influenced by medical care utilization, for example as a result of visiting a doctor, which leads to the problem of endogeneity between medical care utilization and PHB.
Authors [
29] note that visits to GP can significantly influence a patient’s lifestyle by identifying unhealthy behavior. In another similar context, researchers conclude that after allowing for endogeneity, an increase in visits to a GP has a significant positive effect on self-reported health for individuals [
30]. However, in most studies which analyze the involvement of PHB in explaining frequency of visits to the doctor, the potential endogeneity of PHB factors was not controlled for.
We addressed this problem to avoid biased estimates in statistical modeling by using the instrumental variables method. Regarding the influence of disadvantage factors, we expected to find pro-poor inequality in visits to GP and pro-rich inequality to visits to SD, as has been previously reported in many studies of OECD countries that provide universal health insurance [
31,
32]. We also expected to identify additional characteristics which can influence positively or negatively on probabilities of visiting GP or SD, and to find evidence for the above-mentioned statements that the motives for engaging in PHB may be weaker for disadvantaged populations.
Results
The mean values of IndPrev were calculated for various groups of disadvantaged populations.
For age < 60, IndPrev = 0.104 ± 0.002, for older population of age ≥ 60, IndPrev = 0.384 ± 0.009.
For low SES, IndPrev = 0.116 ± 0.004, for middle SES IndPrev = 0.154 ± 0.005, and for high SES IndPrev = 0.171 ± 0.005.
For location in the periphery, IndPrev = 0.133 ± 0.007, in an intermediate location IndPrev = 0.130 ± 0.005, and in the center IndPrev = 0.159 ± 0.004.
For disadvantaged populations with low SES, the value of IndPrev is lower than for populations with middle or high SES, and for those living in peripheral districts, IndPrev is lower than for those living in the center of the country. For older populations, the IndPrev value is some 3.7 times greater than for those under 60 years of age. All of these comparisons are highly significant.
The ordered probit models (Tables
3 and
4) were estimated using IBM Statistics SPSS 20 software (Additional file
1), with the partial marginal effects calculated using MS Excel worksheets (Tables
5 and
6).
Table 3
Estimates, their 95% confidence intervals and significance for the ordered probit model “visits to the GP”
[VisGP = 0] | 1.325 ± 0.050 | 0.000 | 1.344 ± 0.050 | 0.000 | 1.489 ± 0.054 | 0.000 |
[VisGP = 1] | 2.093 ± 0.051 | 0.000 | 2.114 ± 0.051 | 0.000 | 2.261 ± 0.056 | 0.000 |
IndPrev
| | | 0.345 ± 0.041 | 0.000 | 3.037 ± 0.399 | 0.000 |
SES
| -0.065 ± 0.013 | 0.000 | -0.074 ± 0.013 | 0.000 | -0.135 ± 0.016 | 0.000 |
Loc
| 0.002 ± 0.013 | 0.889 | 0.003 ± 0.013 | 0.841 | 0.007 ± 0.014 | 0.605 |
Age
| 0.414 ± 0.027 | 0.000 | 0.342 ± 0.028 | 0.000 | -0.222 ± 0.087 | 0.011 |
Gen
| 0.093 ± 0.019 | 0.000 | 0.083 ± 0.019 | 0.000 | -0.021 ± 0.021 | 0.323 |
Chron
| 0.529 ± 0.024 | 0.000 | 0.499 ± 0.025 | 0.000 | 0.259 ± 0.043 | 0.000 |
SupIns
| 0.096 ± 0.024 | 0.000 | 0.081 ± 0.024 | 0.001 | -0.022 ± 0.028 | 0.433 |
Mar
| 0.016 ± 0.020 | 0.430 | 0.032 ± 0.020 | 0.107 | 0.142 ± 0.026 | 0.000 |
Rel
| -0.019 ± 0.028 | 0.487 | -0.014 ± 0.028 | 0.610 | 0.023 ± 0.028 | 0.419 |
Residuals from the 1st stage | | | | | -2.727 ± 0.402 | 0.000 |
Table 4
Estimates, their 95% confidence intervals and significance for the ordered probit model “visits to the SD”
[VisSD = 0] | 1.924 ± 0.065 | 0.000 | 1.939 ± 0.065 | 0.000 |
[VisSD = 1] | 2.609 ± 0.067 | 0.000 | 2.624 ± 0.067 | 0.000 |
IndPrev
| | | 0.265 ± 0.052 | 0.000 |
SES
| 0.064 ± 0.017 | 0.000 | 0.057 ± 0.017 | 0.001 |
Loc
| 0.028 ± 0.018 | 0.116 | 0.030 ± 0.018 | 0.093 |
Age
| 0.247 ± 0.033 | 0.000 | 0.186 ± 0.035 | 0.000 |
Gen
| 0.188 ± 0.024 | 0.000 | 0.185 ± 0.024 | 0.000 |
Chron
| 0.439 ± 0.030 | 0.000 | 0.415 ± 0.030 | 0.000 |
SupIns
| 0.062 ± 0.032 | 0.048 | 0.050 ± 0.032 | 0.111 |
Mar
| -0.239 ± 0.025 | 0.000 | -0.231 ± 0.025 | 0.000 |
Rel
| -0.080 ± 0.038 | 0.036 | -0.074 ± 0.038 | 0.052 |
Table 5
Partial marginal effects for the ordered probit model “visits to GP”
IndPrev
| -0.588 | 0.416 | 0.173 |
SES
| 0.026 | -0.019 | -0.008 |
Loc
| -0.001 | 0.001 | 0.000 |
Age
| 0.043 | -0.030 | -0.013 |
Gen
| -0.004 | 0.003 | 0.001 |
Chron
| -0.050 | 0.036 | 0.015 |
SupIns
| 0.004 | -0.003 | -0.001 |
Mar
| -0.028 | 0.019 | 0.008 |
Rel
| -0.004 | 0.003 | 0.001 |
Table 6
Partial marginal effects for the ordered probit model “visits to SD”
IndPrev
| -0.028 | 0.020 | 0.007 |
SES
| -0.006 | 0.004 | 0.002 |
Loc
| -0.003 | 0.002 | 0.001 |
Age
| -0.019 | 0.014 | 0.005 |
Gen
| -0.019 | 0.014 | 0.005 |
Chron
| -0.043 | 0.032 | 0.011 |
SupIns
| -0.005 | 0.004 | 0.001 |
Mar
| 0.024 | -0.018 | -0.006 |
Rel
| 0.008 | -0.006 | -0.002 |
Three versions of the visits to GP model were estimated: (a) without IndPrev, (b) with IndPrev and without IV, and (c) with IndPrev and with IV. The variable IndPrev was highly significant in versions (b) and (c), and for the latter, the residuals estimate from the 1st stage was significant. Therefore, the possibility of exogeneity for IndPrev was rejected and version (a) was compared with the version (c).
For the variable
SES, the negative sign of the estimate for both versions (a) and (c) showed that higher values of this variable result in an increased probability that
VisGP = 0 and a decline in the probability that
VisGP = 2. Exactly the opposite follows from the positive signs of the estimates for the dummies
Chron (a 1 value indicating presence of at least one chronic disease) and
Mar (a 1 value indicating unmarried status). For all three variables,
SES,
Chron, and
Mar, the estimates for version (c) were significantly different from those in (a), as their confidence intervals show. The dummy
Age (if 1 then age ≥ 60) estimate changed the sign – its negative sign in version (c) shows that in older populations the probability of
VisGP = 0 increases and the probability of
VisGP = 2 declines. The estimates for
Gen and
SupIns (gender and supplemental health insurance) became non-significant in the version (c), and for
Rel – non-significant in all versions (Table
3).
Two versions of the visits to SD model were estimated: (a) without
IndPrev, and (b), with
IndPrev (instrumentation was not employed in this model). Like in the visits to GP model, the added variable
IndPrev was highly significant in version (b). The estimates for all other variables were not significantly different from those in version (a), as seen from an examination of their confidence intervals. Most estimates were significant at a 5% level, except for
Loc and
SupIns in version (b) (Table
4).
Examination of the partial marginal effects for both models garnered additional information on changes in the probabilities of the subsample level dependent variables when the number of visits equaled 0, 1, or more than 1. For the model for visits to the GP (Table
5), the signs of the effects for the subsample
VisGP = 0 were the opposite of the signs of the estimates for the same version of the model (Table
3, the next to last column). For example, for the variable
SES, as mentioned previously, the negative sign of the estimate (-0.135) shows that an increase in this variable results in an increased probability that
VisGP = 0. In line with this, the values of the partial marginal effects (0.026, -0.019, -0.008) show that a one unit increase in
SES results in increasing the probability that
VisGP = 0 in 2.6%, in decreasing the probability that
VisGP = 1 in 1.9%, and in decreasing the probability that
VisGP = 2 in 0.8%. The sum of the marginal effects of
SES calculated for different values of
VisGP equals zero (allowing for rounding error). This property holds for each explanatory variable (Appendix). For
IndPrev, a 1% increase in this index results in increasing the probability that the person visited GP in the last year, in 0.6% approximately (the partial marginal effect is -0.588 in Table
5).
In a similar manner, the partial marginal effects (Table
6) were analyzed and compared to the estimates (Table
4) for the model of visits to SD. The effect of
IndPrev on visits to SD (Table
6) is much less than the effect of the same variable on visits to the GP (Table
5).
Discussion
This study examined factors explaining the frequency of visits to a doctor for citizens of a country with universal health insurance. The study used data from a national health care survey, and indices calculated from its data, for ordered probit modeling of factors explaining visits to both GP and SD doctors. For both versions of the model used, the same explanatory variables were explored twice, once by taking into account the respondents’ PHB as expressed by the index IndPrev, and once by leaving IndPrev out of the model. This variable was highly significant in both versions, but the influence of its addition on the model estimates was found to be different: there were important changes for visits to GP (in values of the estimates, in their sign, and in their statistical significance) by adding IndPrev, and only slight changes for visits to SD.
In the version of visits to GP, the instrumental variables – housing density (Dens) and father's continent of birth (FatherCont) – were selected to approach the problem of endogeneity between visits to GP and IndPrev. By using the 2SRI method, the exogeneity hypothesis for IndPrev was rejected and the instrumentation was accepted as plausible.
The following variables were identified as significant for explaining frequency of visits to the doctor:
-
Preventive health behavior (PHB) – positive influence of more health preventive behaviors on the probability to visit GP and SD (henceforth “positive” or accordingly “negative”);
-
Socio-economic status (SES) – negative for visits to GP (pro-poor) and positive for visits to SD (pro-rich inequality);
-
Location, gender, and supplemental health insurance, for visits to SD only – positive for central districts (9% level of significance), for females, and for those who have supplemental insurance (5-11% level of significance);
-
Age – age of 60 or greater, negative for visits to GP and positive for visits to SD;
-
One chronic disease at least – positive; and
-
Marital status – if married, negative for visits to GP and positive for visits to SD.
Many findings in this article are in line with previous studies which researched public health under a universal health care system. Our study reveals that the probability to visit GP and SD is positively influenced by PHB which is consistent with results [
44,
29].
Furthermore, we demonstrate pro-poor inequality in visits to GP and pro-rich inequality in visits to SD that are concordant with the findings of other studies conducted in OECD countries where universal health insurance is provided [
31,
32]. In Canada, patients of lower SES were found to have had significantly more primary care visits, while the differences in utilization of specialty services were less pronounced and often not statistically significant [
45]. In Israel, the income-related inequality in primary care is pro-poor, and in secondary physicians’ services is pro-rich [
46].
Our results show the positive influence of location in central districts on the frequency of visits to SD. Similar results were found in [
16], where the authors note that medical care becomes more accessible in areas with higher physician availability, leading to high levels of use.
This study identified age as a factor that can influence the probability of visiting a GP or SD in opposite directions, ceteris paribus, particularly when controlled for PHB and at least one chronic disease factors. This conclusion is similar to the report [
47] where it was concluded that given the high morbidity burden between elder populations, higher use of specialist physicians is found but not higher use of primary care physicians.
Our result that men visited SD less than women agrees with the studies [
48] where it was found that in the USA men are less likely to visit doctors’ offices than women, and [
49] where older females in 11 European studied countries evidenced higher levels of health care usage than males.
Our result that smoking was not related to other factors of PHB (particularly physical activity), made by looking for relevant variables for their inclusion in
IndPrev, is in agreement with the results in [
50] where physical activity was found unrelated to smoking for older persons in Israel.
The main limitations of this study are found in the choice of explanatory variables, because of the absence of data on health status and income of respondents in the National Health Survey questionnaire. These issues were addressed by using the “one chronic disease at least” question as a proxy for health status, and by using occupation and education level as proxies for income.
Appendix
Calculation of partial marginal effects
Parameter estimates from the ordered probit model (3) can be transformed to estimates of the partial marginal effects, which show the change in predicted probabilities of specific values of the dependent variable associated with changes in the explanatory variables. In the ordered probit model, a positive (negative) value of β
k
means that higher values of x
k
increase (decrease) the likelihood of higher values of the dependent variable VisGP (and VisSD, accordingly). Unlike the OLS regression, coefficients in the ordered probit model do not represent marginal changes in the dependent variable, given minor changes in the independent variables. The influence of a change in x
k
on the probability p that VisGP or VisSD receive one of their possible values (0, 1, or 2) can be examined by taking the partial derivative of this probability, with respect to x
k
("partial marginal effect" of x
k
).
For the model with
VisGP, and for each of continuous
x
k
the partial marginal effects
ME0,k (when
VisGP = 0),
ME1,k (when
VisGP = 1),
ME2,k (when
VisGP = 2) is calculated as follows:
(6)
The α0,α1 are defined in (4), and φ denotes standard normal density function.
For discrete variables, for example for a binary (0, 1) variable x
k
, the effect of the zero-to-one discrete change for each value j of VisGP (j = 0, 1, or 2) is calculated simply as a difference between the probabilities of two events: VisGP = j for x
k
= 1 and for x
k
= 0. The effects of discrete changes for variables (1, 2, 3) are calculated in a similar manner.
In addition, for each of the explanatory variables the sum of its marginal effects calculated for different values of VisGP, equals 0. This follows immediately from (6) and from the described procedure for calculating marginal effects of the discrete variables.
For the model with
VisSD, the partial marginal effects are calculated in a similar manner [
51].
Gregory Yom Din, PhD, is a coordinator of the course of applied econometrics for MBA at the Open University of Israel and a researcher at the Faculty of Exact Sciences, Tel-Aviv University. His current research interests include econometric modelling of health care systems, particularly for disadvantaged populations. Zinaida Zugman, PhD, is a former researcher at Golan Research Institute, Katzrin, Israel and an independent consultant, with a special focus on development of mathematical models and simulation models for decision-making. Alla Khashper recently joined the department of Diagnostic Imaging at Soroka University Medical Center, Beer Sheva, Israel after completing her fellowship at McGill University, Montreal, Quebec. She is a lecturer at the medical school of Ben-Gurion University of the Negev, Beer-Sheva, Israel for local and international students.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
GYD initiated the study and took the lead in its planning, implementation, and writing the article. ZZ and AK participated in implementation of the study and writing the article. All of the authors have reviewed and approved the final manuscript.